Note to self: Seasonality is when a pattern repeats in a given week or year. Seasonal data is not stationary but can be made so. This is through taking z_t = y_t - y_(t-365). Similar process to de-trending data. Note that seasonality is not the same as cycles. Cycles are usually over the course of multiple years!
Thanks for these videos; they're helpful. But I think one big thing is missing which is what we're actually trying to model. We've seen what's bad (trend and seasonality) but I think it would be good to have an example of what's good. What would an ideal stationary time series look like? And how would the model come out?
i think Z(t) = Y(t) - Y(t-365). This would also track with the example of 2015 that we can't use. cuz if we use Z(t) = Y(t+365) - Y(t), we can totally replace it with 2015. Also the differenciation to remove seasonality (I in ARIMA) is given by : Y'(t) = Y(t) - Y(t-1).
When he says "within a year" he just meant that the pattern doesn't have a period longer than a year (which would then be considered a cycle). Therefore, as he said at the beginning, the cycle could be a year but also a month, a week or even a day
Is it so that ADF & KPSS tests check only for trend-stationarity & therefore, it is possible for ADF & KPSS tests to show that the time series is trend-stationary but still has seasonality in it? If yes, then what are tests to check seasonality-stationary?
where there is the start of 2015/2016/2017.. and so on, the curve should be lower and not higher because of the winter period.the start of the year,2015,is the 1/01/yyyy and not the summer
so in order to know whether a pattern is seasonal, you would need at least two years of data right? If you have one year of data and no matter how smooth sign wavy it looks, you cant call it seasonal? I mean one cant look at data for a year and say this pattern looks seasonal. correct?
Yea I was thinking 4 years, like search data for the Olympics for example. But I think even so, we should be able to apply the same concept to remove the effect of these repeating patterns on our time series data. So long as we have sufficient cycles of data of course (in your example, a few decades)
Note to self: Seasonality is when a pattern repeats in a given week or year. Seasonal data is not stationary but can be made so. This is through taking z_t = y_t - y_(t-365). Similar process to de-trending data. Note that seasonality is not the same as cycles. Cycles are usually over the course of multiple years!
Good introductory/reference explanation. Thank you!
Glad it was helpful!
Thank you so much.
Thanks for these videos; they're helpful. But I think one big thing is missing which is what we're actually trying to model. We've seen what's bad (trend and seasonality) but I think it would be good to have an example of what's good. What would an ideal stationary time series look like? And how would the model come out?
Amazing work Ritvik
Amazing, simple explanation. Thank you.
i think Z(t) = Y(t) - Y(t-365). This would also track with the example of 2015 that we can't use. cuz if we use Z(t) = Y(t+365) - Y(t), we can totally replace it with 2015. Also the differenciation to remove seasonality (I in ARIMA) is given by : Y'(t) = Y(t) - Y(t-1).
Fantastic explanation
This series is so good. Thank you so much!
Very easy to understand. Thank you so much
awsome explanation!
hi, can you sort the videos according to series. Ironically, i cant figure out which time series video i should watch first.
Just sorted them! Thanks for the reminder 😊
Dude, thank you so much.
Could you make a video explaining a vector autoregression model?
ua-cam.com/video/UQQHSbeIaB0/v-deo.html
thank u sir
Nice Insights!
thanks! always happy to help out :)
Sir please a video on Johnsens cointegration.. Please.. Detailed one.. You are amazing sir ❤️❤️❤️❤️
From the video it seems seasonality means yearly repeating patterns but what if the patterns repeat every week / month/ 3 moths /6 months etc
The main importance is simply that it is a nonrandom component that can be addressed with a model. The yearly unit is truly arbitrary.
When he says "within a year" he just meant that the pattern doesn't have a period longer than a year (which would then be considered a cycle). Therefore, as he said at the beginning, the cycle could be a year but also a month, a week or even a day
Is it so that ADF & KPSS tests check only for trend-stationarity & therefore, it is possible for ADF & KPSS tests to show that the time series is trend-stationary but still has seasonality in it?
If yes, then what are tests to check seasonality-stationary?
Can we use ARIMA model for seasonality predictions like electricity consumption predictions, or we should remove seasonality before applying ARIMA
where there is the start of 2015/2016/2017.. and so on, the curve should be lower and not higher because of the winter period.the start of the year,2015,is the 1/01/yyyy and not the summer
How about a video on Expectations
so in order to know whether a pattern is seasonal, you would need at least two years of data right? If you have one year of data and no matter how smooth sign wavy it looks, you cant call it seasonal?
I mean one cant look at data for a year and say this pattern looks seasonal. correct?
What if there is a fixed length cycle with period greater than a year, a decade for instance?
Yea I was thinking 4 years, like search data for the Olympics for example. But I think even so, we should be able to apply the same concept to remove the effect of these repeating patterns on our time series data. So long as we have sufficient cycles of data of course (in your example, a few decades)
How to find seasonality or determine m value if we have 3 or 6 months of data ?
You simply don't have enough data in that case
Cool !!!
wr
Y